Deep Research Orchestrator

Ask a hard question. Watch the search spend its budget.

Simulated metrics — no real experiments

Ask a hard question. Watch the search spend its budget.

Type a research question. The search grows a tree of experiment ideas from it. Every number is simulated, so the same question always grows the same tree.

Runs on a $30 pretend budget. Same question, same seed, same tree.

How the search works

Four steps, repeated until the search reaches the ideas cap or the pretend budget runs out, whichever comes first. No web research happens here: the “experiments” are simulated numbers from a fixed formula.

  1. Ask

    Your question becomes the seed of the tree: the idea everything else grows from.

  2. Grow

    The search makes five small edits per idea: a new sparsity, a new learning rate.

  3. Score

    Each idea is scored: 0.4 × novelty + 0.5 × quality + 0.1 × budget left.

  4. Pick

    The best score runs first. The winning chain of ideas is the answer.

Search settings
$30

The search stops when spend reaches the budget. One experiment may push it slightly over.

24

Checked each time the search picks its next idea. Every pick adds five, so the count can land a little above this number.

Change it for a different tree. Keep it to replay the same one.

What the first five ideas will be

  1. sparsity top-4
  2. sparsity top-8
  3. sparsity top-16
  4. lr=0.0003
  5. lr=0.001

Ideas come from small fixed edits to one knob at a time — no model plans them. Ideas are picked best-first: 0.4 × novelty + 0.5 × quality + 0.1 × budget left. About one in ten fails, and failed ideas are never picked again.

Search graph
ok FAIL frontier seedbest branch

No search yet. Type a question above, then press Run the search.

Ideas tried

–

Budget spent (simulated)

–

Failed ideas (simulated)

–

Best loss (simulated)

–

Search log

The trace reads like the original main.py output. Costs and losses are simulated; nothing really runs.

Best chain of ideas

The highest-scoring chain of small edits, in the order the search picked them.

The winning chain appears here once a search finishes.